基于改进YOLOv4算法的轮毂表面缺陷检测

Surface Defect Detection of Wheel Hub Based on Improved YOLOv4 Algorithm

汽车轮毂加工过程中产生的表面缺陷严重影响整车的美观性及服役性能,针对人工检测效率低、漏检率高的问题,提出一种基于改进YOLOv4算法的轮毂表面缺陷检测方法。构建了轮毂缺陷数据集,其包含6种表面缺陷,由2346张4928×3264 pixel的图像组成;采用K-means方法进行先验框聚类,并针对YOLOv4算法在纤维、粘铝等小尺度缺陷上检测精度不足问题,在原网络Neck部分引入细化U型网络模块(TUM)和注意力机制,用于增强有效特征并抑制无效特征,强化多尺度特征提取与融合,改善特征处理过程中可能存在的小目标信息丢失问题;基于该数据集,训练并测试不同算法的缺陷检测性能并验证改进模块的有效性。结果表明,该方法大幅提升了粘铝等小尺寸缺陷的检测能力,缺陷检测平均精度达到85.8%,与多种算法相比较检测精度最高。

The surface defects produced during the processing of automobile wheels seriously affect the aesthetics and service performance of the whole vehicle.Aiming at the problems of low manual defect detection efficiency and high missed detection rate at this stage,a method for detecting wheel surface defects based on the improved YOLOv4 algorithm is proposed.Constructed a wheel defect data set,including six types of surface defects,composed of 2346 images of 4928×3264 pixel,using K-means method for priori box clustering,and focusing on YOLOv4 algorithm on small-scale defects such as fiber and sticky aluminum insufficient detection accuracy.In the Neck part of the original network,a thinned U-shaped network module(TUM)and attention mechanism are introduced to enhance effective features and suppress invalid features,strengthen multi-scale feature extraction and fusion,and improve the possibility of feature processing small target information loss problem;based on self-built data sets,training and testing the defect detection performance of different algorithms and verify the effectiveness of the improved modules,the results show that the average defect detection accuracy of the method reaches 85.8%,and it greatly improves small-size defects such as aluminum sticking.The detection ability of,the detection accuracy is the highest among many comparison algorithms.